Bilinear mode in topi::nn::upsampling not working

[ke1337]

To repro, use following changes in test_topi_upsampling.py:

--- a/topi/tests/python/test_topi_upsampling.py
+++ b/topi/tests/python/test_topi_upsampling.py
@@ -22,7 +22,7 @@ def verify_upsampling(batch, in_channel, in_height, in_width, scale, layout='NCH
         raise NotImplementedError(
             'Layout not supported {} '.format(layout))

-    B = topi.nn.upsampling(A, scale, layout=layout)
+    B = topi.nn.upsampling(A, scale, layout=layout, method='BILINEAR')

     b_np = topi.testing.upsampling_python(a_np, scale, layout)

[masahi]

cc @srkreddy1238

[srkreddy1238]

@KeDengMS

NCH layout ???
Can you conform the error log ?

[ke1337]

Yes NCHW layout, but I believe it exists in NHWC as well. The error is from a crash in dereferencing nullptr in here. The problem is that shape is a Mul node with two IntImm as input from the python API, and as_const_int returns nullptr for non-const node.
It seems const folding should be applied, and moreover the bilinear interpolation code path needs to be tested.

[srkreddy1238]

@KeDengMS a bit confused.

Is the issue about "Layout Not Supported" or "Crash " ?

Can you share complete log or the sample code?

[ke1337]

I think it's nothing about Layout, but more for BILINEAR mode. The repro is to change the python test to use BILINEAR mode: B = topi.nn.upsampling(A, scale, layout=layout, method='BILINEAR'), then you'll find the crash.

[masahi]

But the log says NotImplementedError('Layout not supported {} '.format(layout)). What is the "layout" parameter you passed in? It should be "NCHW" or "NHWC".

[ke1337]

The repro I show above is from git diff, not the log. The testcase tests both NCHW and NHWC

[masahi]

@KeDengMS I can repro your crash. The fix is to add util.simplify to https://github.com/dmlc/tvm/blob/master/topi/python/topi/nn/upsampling.py#L34
This is equivalent to const folding you mentioned.

I can send a PR to fix this and update the upsampling test.

[masahi]

@srkreddy1238 I made a branch to fix this issue here. But some bilinear upsampling tests are failing. Can you have a look?

[srkreddy1238]

Thanks @KeDengMS for bringing up the issue and thanks @masahi for finding the root cause and fix.
I have added the final piece (atol=1e-5 allowance) at #1708 for your review comment.

[ke1337]

Thanks for the fix!